Multi-Label Cytology Single-Cell Image Dataset for Nuclear Feature Classification

Access to this dataset is restricted. To request access, please contact the corresponding author at [[e-mail]] stating your research purpose. Users are required to cite both the associated paper and this dataset in any resulting publications. This dataset contains 3,419 expert-annotated single-cell images derived from thyroid fine-needle aspiration cytology slides, prepared at the 108 Military Central Hospital, Vietnam. Each image is labeled with nine clinically relevant nuclear morphological features (e.g., nuclear elongation, grooves, inclusions, enlargement, overlapping), following the Bethesda System for Reporting Thyroid Cytopathology. The annotation was performed by a team of seven board-certified cytopathologists with 5–20 years of experience. Each cell was independently reviewed by two experts, with disagreements adjudicated by a third senior pathologist, and final labels established by group consensus. To standardize preprocessing, cells were cropped from original microscopy images, centered on a 224×224 white background, and upscaled from 20× magnification. The resulting dataset provides a clean and uniform input format for deep learning models while preserving key cytological features. The dataset is inherently multi-label and imbalanced, reflecting real-world cytopathology where some features (e.g., nuclear grooves) are frequent while others (e.g., intranuclear inclusions) are rare. This makes it particularly suitable for benchmarking methods addressing class imbalance, label correlations, calibration, and interpretability in multi-label learning. This resource is intended for research in computational pathology, deep learning, and medical image analysis, and can be used to train, validate, and benchmark algorithms for thyroid cancer diagnosis.